Class-wise confidence-aware active learning for laparoscopic images segmentation.

Area of Science:

• Medical Image Analysis
• Computer-Assisted Surgery
• Machine Learning

Background:

• Segmentation is crucial for computer-assisted surgery (CAS) systems, defining organ shapes and instrument locations.
• A major barrier to powerful segmentation approaches is the need for extensive annotated data.
• Active learning (AL) aims to reduce annotation workload by selecting informative samples, but often overlooks low-frequency classes common in laparoscopic videos.

Purpose of the Study:

• To address the poor segmentation performance caused by the failure of previous AL methods to select low-frequency classes in laparoscopic videos.
• To improve AL strategies by exploring the use of unselected data and focusing on underrepresented classes.
• To enhance the practical application of powerful segmentation techniques in CAS by reducing annotation requirements.

Main Methods:

• Proposed a class-wise confidence bank to store and update confidence scores for each class.
• Developed a new acquisition function leveraging the confidence bank to guide sample selection.
• Utilized confidence scores in conjunction with a class-wise data mixture method to exploit unlabeled data without additional annotation.

Main Results:

• The proposed method demonstrated superior performance compared to previous AL studies on the CholecSeg8k dataset, achieving approximately [Formula: see text] improvement.
• Significant performance gains were observed specifically for low-appearing frequency classes.
• Experiments on the RobSeg2017 dataset, with varying annotation budgets, confirmed the effectiveness of the proposed approach.

Conclusions:

• A class-wise confidence score effectively enhances the acquisition function for active learning.
• Exploiting unlabeled data with the proposed confidence score leads to substantial improvements over existing methods.
• The method successfully boosts segmentation performance for low-frequency classes, a critical advancement for surgical video analysis.